OpenClaw 2.0 Explained: What's New in v2026.8.1

OpenClaw 2.0 — version 2026.8.1 — is the biggest release in the platform's history, and the first one that behaves like a traditional major version. Before this, OpenClaw shipped 106 releases in 230 days: a firehose of daily improvements that made the project famous and occasionally made upgrading exciting in the wrong way. For 2.0, the team paused that cadence for almost two months to harden the foundation, then shipped everything at once.

The numbers are unusual for any open-source project:

• 933 contributors, 569 of them contributing for the first time
• More than 16,000 merged pull requests
• Changes across installation, agents, plugins, credentials, browser control, messaging, automation, memory, and the native apps

The version number still follows OpenClaw's calendar scheme (2026.8.1 = August 2026), but the project brands this one 2.0 because it changes the defaults you live with every day: how sessions get permissions, how credentials are handled, how fast the Gateway starts, and how much of the platform you can run without touching a config file.

The full technical changelog is in the official release notes. This post is the practical read: what actually changed, what breaks, and what it means if you are building or buying agents on top of OpenClaw.

The release notes lead with eight user-facing features. In rough order of daily impact:

• Conversation search. Search visible text across past conversations by exact words or phrases, and jump back into the surrounding messages. Sounds small; changes everything once your agent has months of history.
• Sessions beyond your Gateway. Run work on paired devices or cloud workers, and move a session's workspace between machines. One person can start a task and a teammate can pick it up with full context.
• Durable progress cards. A session's plan and status persist through reloads across web, macOS, iOS, and Android — you can actually follow long-running agent work instead of staring at a spinner.
• Structured questions. When the agent needs input, it asks through native cards and buttons on your channel instead of free-text guessing, with an explicit skip path.
• Interactive results. Widgets render in chat, can be pinned to dashboards, and only get the specific actions and network origins you grant.
• Private credential requests. The agent can ask for a secret through a masked prompt — the value never enters the chat history or the model context.
• Approve recurring work once. Automations request a permission grant one time; if the operation later changes, the approval is invalidated and requested fresh. No silent authority creep.
• Richer media. Audio and video stay attached to conversations across uploads, replies, and reloads, with native playback controls and video upload on iOS and Android.

If you run an agent on Telegram or WhatsApp, the combination of structured questions and durable progress cards is the one you will feel first — long tasks stop feeling like a black box.

Strip away the feature list and OpenClaw 2.0 is a security release. That tracks with what the platform has become: an agent with access to your files, browser sessions, messaging accounts, and cloud workers is a very attractive target, and 2026 has supplied plenty of cautionary tales.

What changed:

• Session permission modes. Every session now runs in an explicit permission mode, with filesystem access anchored to the recorded workspace or worktree — an agent can no longer wander the disk by default.
• Plugin trust review. Before an external plugin is installed or enabled, OpenClaw shows its capabilities, source, version, and exact artifacts. Decline an update and the previous version keeps running unchanged.
• A real credential layer. Secrets live in a shared credential store as write-only values, scoped to teams, with network egress bound to declared hosts. An optional 1Password broker adds per-secret approval with a value-free audit trail.
• Model allowlists. Operators can restrict which models any agent may call, per agent, with provider wildcards.
• Configuration change history. Config changes are recorded with writer labels and sensitive values redacted — you can see who changed what, and when.

If you have been through the OpenClaw Masterclass, this is the direction its Permissions & Sandboxing and Security & Incident Response sections have argued for all along: give the agent the narrowest workable authority and make every escalation explicit. In 2.0, those practices stop being discipline and start being defaults.

Channel work in 2.0 is substantial. What matters most on the three channels OpenClaw deployments use most:

Telegram got the flashiest upgrade: a full Mini App dashboard. Send /dashboard to your bot and the entire Control UI opens inside Telegram — no separate browser tab. Messages also got richer through Bot API 10.2: native tables, headings, checklists, and formulas render properly instead of as plain text, and live location updates now reach the agent as structured events.

Discord agents can now join voice rooms (and are told when participants join or leave), discover custom emoji, and render interactive widgets inside Discord Activities. Presence triggers are new too: an agent can optionally wake when a specific participant comes online, with scoped audiences and cooldowns.

WhatsApp work this cycle was stability-focused: replies now stay attached to the exact direct or group message being answered instead of appearing as detached follow-ups, and the session's auth and lease are held until the Gateway actually stops — fewer surprise re-link dances.

Slack deserves a mention even though it is newer to the roster: 2.0 makes it first-class, with Enterprise Grid support, direct messages, workspace routing, approvals, and native charts and tables in messages.

Still deciding where to deploy? The trade-offs in our Telegram vs Discord vs WhatsApp comparison hold in 2.0 — if anything, Telegram's lead widened with the Mini App dashboard.

Three changes push OpenClaw further toward an agent that improves with use:

• Grounded dreaming is now on by default: a model-backed background process consolidates recent activity into long-term memory, promotes only provenance-qualified material, and logs what it did to a reviewable Dream Diary.
• Automatic self-learning captures strong, reusable lessons from sessions and applies scanner-approved skills by default. The new Skill Workshop consolidates what the agent has learned into curated collections through a recoverable review flow — it holds at most three pending proposals, so it stays reviewable instead of becoming a junk drawer.
• Memory ownership tools landed alongside: inspect which sessions contributed to memory, exclude sources you don't want learned from, and delete identifiable derived memories with openclaw memory forget while keeping source transcripts intact.

For skill authors and buyers, the important non-change: the SKILL.md format is untouched. Skills built for 1.x — including everything on the OpenClaw marketplace — load and run on 2.0 without modification. 2.0 also adds skills-sh: references, which install commit-pinned skill artifacts mirrored through ClawHub with the normal safety checks applied.

The performance work is easy to summarize: the Gateway now starts in about 575 milliseconds, down from roughly 1.6 seconds, and default agent concurrency scales with your CPU (8–16 simultaneous runs on typical hardware). A long reliability pass hardened the SQLite layer — snapshot verification, corruption quarantine that survives further damage, protection against write-ahead-log split-brain — which matters more than it sounds for an always-on agent.

Onboarding was rebuilt around detection: the first-run flow finds what you already have — a Claude or ChatGPT subscription, API keys, local models — and offers it as structured choices instead of a config file. It can even migrate memory from existing Claude Code, Codex, and Hermes installs.

On the model side, 2.0 adds live model discovery (pulling what your providers actually offer instead of a baked-in catalog) and ships discovery for new entries like Qwen 3.8 Max and Flash, DeepSeek V4 Flash Vision, NVIDIA Nemotron 3.5 Lightning, and GLM-5.3. Anthropic users can opt into beta server-side context compaction while keeping the full transcript locally.

None of this changes the model guidance for OpenClaw skills: pick a frontier model for judgment-heavy skills and a fast, cheap model for high-volume ones.

For a release this size, the breaking-change list is remarkably short. Three items:

1. OpenProse is retired. The bundled plugin and its /prose command are removed. Your .prose files are untouched; openclaw doctor --fix cleans up the configuration, and the upstream project documents an Agent Skill migration.
2. OpenAI routes were unified. Model references under codex/* and openai-codex/* migrate to openai/* — provider config, stored sessions, and automations included. Again: openclaw doctor --fix does the rewrite for you.
3. Plugin SDK subpaths are deprecated. External plugin authors must move five legacy subpath imports to the focused openclaw/plugin-sdk entry points by September 1, 2026. If you only use plugins, nothing changes until your plugin's author updates it.

Everything else — skills, channel configs, automations, memory — carries over. We wrote a separate step-by-step update guide covering backups, the doctor migrations, and exactly what to verify afterward.

Less than you might expect. OpenClaw 2.0 changes defaults and adds surface area, but the architecture underneath — a Gateway you run, skills that define what your agent can do, channels that decide where it talks to you — is exactly the same mental model.

That is good news if you are learning it now: the OpenClaw Masterclass teaches that model in the order 2.0 rewards — foundations and architecture, install and configure, permissions and sandboxing, skills and automation, then security and incident response. The 2.0 changes slot into that structure rather than replacing it: permission modes are the sandboxing chapter formalized into product, the credential store is the security chapter's practices shipped as defaults, and the Skill Workshop builds on the same SKILL.md format the skill chapters teach.

Every tier of the Masterclass ($49.99 and up) includes lifetime updates, so the curriculum tracks the platform as it moves — that is the point of buying a course once instead of subscribing to content that ages. And if you mostly want the outcome rather than the theory, the marketplace skills run identically on 2.0.